From vital sign trajectories to data-driven targets: defining exploratory blood gas ranges in sepsis-associated thrombocytopenia.

Background: The early in-intensive care unit (ICU) phase is critical for sepsis-associated thrombocytopenia (SATP) patients, yet the prognostic value of their initial physiological trajectory remains underexplored. We aimed to identify distinct subgroups based on vital sign trajectories following ICU admission and to investigate their differential outcomes and subsequent blood gas management needs. Methods: This retrospective study utilized the MIMIC-IV database. Adults with SATP were included. Group-based multi-trajectory modeling (GBMTM) was applied to hourly vital signs (including heart rate, blood pressure, respiratory rate, and SpO₂) from the first 12 h of ICU stay to identify subgroups. Mortality risk was assessed using Cox regression, with the lowest-risk cluster as the reference. Within the identified high-risk sub-phenotype, the nonlinear relationships between blood gas ranges and ICU mortality were analyzed with restricted cubic splines (RCS). Finally, multivariable partial dependence plots (PDP) were employed to quantify the optimal ranges for blood gas parameters, defined as those associated with the lowest ranges of predicted mortality risk for this subgroup. Results: The analysis of initial 12-h physiological trajectories classified patients into three subgroups: Cluster 1 (characterized by elevated blood pressure), Cluster 2 (marked by high heart rates and respiratory rates with low SpO₂), and Cluster 3 (low blood pressure with high SpO₂). Cluster 2 was identified as the high-risk subgroup, demonstrating significantly increased mortality risks compared with Cluster 3: ICU mortality (HR = 1.40; 95% CI: 1.13-1.73), 28-day mortality (HR = 1.56; 95% CI: 1.30-1.88), 90-day mortality (HR = 1.43; 95% CI: 1.21-1.67), and 365-day mortality (HR = 1.33; 95% CI: 1.15-1.54). Within Cluster 2, restricted cubic spline analyses revealed nonlinear relationships between blood gas parameters and ICU mortality. Using partial dependence plot analysis, we identified model-derived ranges of blood gas values associated with the lowest predicted mortality risk, which may serve as exploratory physiological references for this high-risk subgroup: pH 7.32-7.64, PO₂ 25.00-324.32 mmHg, PCO₂ 21.94-53.74 mmHg, lactate 0.6-7.49 mmol/L, base excess -7.47 to 23.00 mEq/L, and total CO₂ 43.47-56.00 mEq/L. These ranges, though broad, reflect the inherent physiological variability during the early ICU phase and should be interpreted as hypothesis-generating parameters rather than strict clinical targets. Conclusion: Early vital sign trajectories during the first 12 h in the ICU effectively stratify SATP patients into prognostic subgroups. For the high-risk subphenotype, we further delineated model-derived physiological ranges of blood gas parameters, creating a "trajectory-to-targets" framework. This approach offers a hypothesis-generating strategy for transitioning from early risk identification to personalized physiological insights in the critical early phase of ICU care.